An effective AI infused demand forecasting application for automotive spare parts industry: a real case from Turkey

Why guessing spare parts demand matters

When a car breaks down, drivers expect the right spare part to be on the shelf. For manufacturers and service shops, keeping enough parts without overfilling warehouses is a constant juggling act. This study looks at how smart use of artificial intelligence can help companies in the automotive spare parts industry predict very irregular demand more accurately, saving money while keeping customers on the road.

The challenge of stop-and-go demand

Spare parts do not sell in a smooth, steady stream. Many items sit for months with no orders, then see sudden bursts of demand. The authors call these patterns intermittent and lumpy demand. In many time periods, demand is zero; when orders do appear, the quantities can swing wildly. Classic forecasting tools that work well for everyday goods struggle in this setting, often leading either to empty shelves or to expensive piles of unused stock. Because spare parts are critical for keeping equipment running, poor forecasts can damage service quality and increase costs.

From old rules to learning from data

To tackle this, the researchers reviewed both long-standing statistical methods and newer data-driven approaches. Traditional techniques known as Croston-based methods were designed specifically for series with lots of zeros, but they have limits when demand becomes highly erratic. Newer machine learning methods such as support vector machines, random forests, and linear regression, and deep learning methods such as multilayer perceptrons, recurrent neural networks, and long short-term memory networks, can automatically learn complex patterns from data. However, each method has its own strengths and weaknesses, and no single model is best in every situation.

Real data from a spare parts maker

The team worked with a Turkish manufacturer that produces electric drive system components for vehicles. They focused on two high-value spare parts that are especially hard to plan: one with intermittent demand and one with strongly lumpy demand. Using 51 months of real monthly sales data, they cleaned and normalized the records, then classified the demand patterns with standard measures of how often demand occurs and how much it varies. The data were split into training and testing sets so that models would be evaluated on unseen periods, mirroring real-world forecasting.

Letting many models vote together

Instead of relying on a single forecasting tool, the authors built a stacking ensemble, which can be thought of as a team of models whose outputs are combined by a simple second-level model. First, several machine learning and deep learning models independently produced forecasts. Their predictions were then fed into a linear regression model that learned how much to trust each one. This stacked model was compared against Croston-based methods and each individual learning method, using error measures that capture typical mistakes and how large they are relative to average demand. The researchers also used graphical summaries and statistical tests to check whether performance differences were meaningful rather than due to chance.

What the smarter forecasts mean for business

The stacked model consistently produced the most accurate and stable forecasts for both intermittent and lumpy demand. Deep learning methods on their own generally outperformed traditional approaches, but combining multiple methods worked best of all. For the partner company, this meant lower safety stocks, fewer obsolete items left on shelves, and a reduced risk of failing to supply customers on time. In simple terms, the study shows that using a carefully designed mix of AI models can turn messy, stop-and-go spare parts demand into predictions that are good enough to support leaner inventories and more reliable service.

Citation: Albayrak Ünal, Ö., Erkayman, B. & Usanmaz, B. An effective AI infused demand forecasting application for automotive spare parts industry: a real case from Turkey. Sci Rep 16, 15661 (2026). https://doi.org/10.1038/s41598-026-44461-0

Keywords: spare parts demand, intermittent demand, ensemble learning, automotive supply chain, demand forecasting